- The paper presents ISM, a self-improving strategy memory that enhances continual mathematical reasoning through verifiable, autonomous schema updates.
- ISM leverages a dual-representation schema and hybrid feature extraction to achieve up to 80.67% accuracy on MATH-Hard while maintaining strict episodic isolation.
- The implementation significantly improves memory efficiency by using 20× fewer schemas than retrieval-based systems, reducing catastrophic forgetting.
Intelligent Schema Memory for Continual Mathematical Reasoning
Introduction
The paper introduces Intelligent Schema Memory (ISM), a memory-augmented, self-evolving system designed for continual mathematical reasoning with a frozen LLM (2606.31191). ISM addresses the persistent challenge of enabling cumulative improvement in reasoning systems where direct parameter updates are not permitted, focusing on the use-case of mathematical problem streams with strict episodic isolation. The system maintains a compact repository of strategy schemas, refined and regulated via autonomous online mechanisms, with every update gated through symbolic verification. The results demonstrate substantial gains in performance and memory efficiency for ISM over multiple strong baselines, establishing new best results under limited-memory continual learning regimes on benchmarks such as MATH-Hard and OlympiadBench.
Methodological Framework
ISM operates under a continual learning protocol with hard episodic resets: after each math problem, the frozen LLM is reinitialized, and only structured, external memory persists. The ISM architecture consists of five principal components:
- Hybrid Feature Extractor: Maps problems into structured feature tuples via a combination of rule-based and lightweight LLM classifiers, integrating operator type, problem structure, heuristic signatures, and semantic embeddings. An agreement scoring mechanism ensures robust extraction by reconciling the two branches.
- Schema Bank: Stores strategy schemas, each partitioned into stable content (strategy description, template, heuristics) and an adaptive feature hook (retrieval-side signals such as operator, structure, heuristics, embedding centroid, and success rate).
- Two-Stage Retrieval: For each new problem, an operator-type filter constrains candidates, followed by a weighted soft scoring over structure, heuristics, quantities, embeddings, and a prior on schema success rate. Retrieval confidence thresholds select between specific, partial, or generic schemas.
- Frozen LLM Solver and Tools: The retrieved schema content is injected into the LLM prompt. Optionally, agentic tool use is triggered for hard cases, activating symbolic verifiers, toolchains, or error-search aids within a bounded turn budget.
- Memory Controller and Self-Improvement: Seven asynchronous mechanisms maintain the schema bank: periodic audit, correction, merging of near-duplicates, pruning, promote/demote by success, reinforce by heuristics, and antipattern extraction from failures. Every update is regulated by symbolic verification.
A distinctive aspect of ISM is its dual representation of schemas, decoupling retrieval adaptation from content stability. All schema evolution—including synthesis in response to repeated failures—is tightly controlled via symbolic certificate gates.
Empirical Evaluation
ISM is rigorously benchmarked on two demanding datasets: MATH-Hard (selected from the hardest problems in MATH) and OlympiadBench (Olympiad-level, multi-step mathematics problems), both processed as 300-episode streams with adversarial domain shifts.
Baselines:
- Vanilla frozen LLM (no external memory)
- Static schema (one immutable template)
- Retrieval-augmented generation (entire episodic memory, embedding retrieval)
- Reflection-based memory (stores per-episode free-text feedback)
- Passive schema memory (no self-improvement or verification mechanisms)
Metrics include global accuracy, domain plasticity (early-block accuracy), stability (late-block accuracy), catastrophic forgetting (accuracy drops across domain revisiting), backward transfer, and schema bank storage size.
Results:
| Benchmark |
System |
Accuracy |
Stability |
Forgetting |
Bank Size |
| MATH-Hard |
ISM |
80.67 |
84.67 |
0.07 |
17 |
|
Passive Schema Memory |
78.67 |
82.0 |
0.10 |
47 |
|
RAG |
57.00 |
60.0 |
0.12 |
300 |
| OlympiadBench |
ISM |
61.67 |
64.00 |
0.03 |
13 |
|
Passive Schema Memory |
59.67 |
62.0 |
0.10 |
91 |
|
RAG |
33.33 |
34.00 |
0.02 |
300 |
ISM surpasses all baselines on accuracy, stability, and catastrophic forgetting while using 64–86% fewer schemas than the best passive baseline and over 20× fewer than retrieval-based systems. The memory footprint remains bounded and shrinks via active maintenance, contrasting sharply with baselines where unused or redundant schemas accumulate linearly with problem stream length.
Additionally, ISM demonstrates greater resilience to domain shifts and maintains strictly positive backward transfer on OlympiadBench, indicating that schema refinement not only prevents forgetting but enables knowledge consolidation across tasks.
Implications and Discussion
The primary implication is that self-evolving, verifiable external strategy memory enables cumulative, trustworthy improvement in frozen LLMs under strict continual learning constraints. ISM's lifecycle mechanisms (audit, merge, prune, reinforce, antipattern extraction), driven by outcome-gated symbolic verification, ensure that only validated, durable knowledge persists. This obviates storage bloat and guards against error reinforcement—a common failing in pure retrieval or reflection-based setups.
From a practical perspective, ISM opens the door to more efficient and reliable reasoning engines in settings where model parameters cannot be retrained or updated (e.g., LLMs deployed in constrained or regulatory contexts). The compact schema bank provides transparency, interpretability, and memory efficiency, which are highly desirable in both academic and deployment scenarios.
Theoretically, the success of ISM suggests a template for integrating active memory management, symmetric learning from successes and failures, and formal verification into the broader landscape of continual learning with LLMs. ISM's design is model-agnostic and could, in principle, be hybridized with substrate-oriented innovations, such as systems with trainable long-term memory modules.
Limitations and Future Directions
Several limitations are acknowledged:
- Results are based on a single stream order; robustness to order variance and multiple seeds remains to be established.
- Per-mechanism ablations are not presented, leaving the contribution of individual self-improvement modules open for further study.
- Comparisons do not extend to search-based or agentic architectures like Tree-of-Thoughts or graph-structured memory.
- Application is currently domain-limited to mathematical problem solving; generalization to proof generation, code, or multimodal reasoning is untested.
- Verification depends on symbolic mathematics, which may not generalize to domains lacking formal correctness or requiring proof-theoretic checking.
Future work should include comprehensive ablation analyses, evaluation across additional domains and LLM sizes, robust multi-seed and stream ordering tests, and integration with richer formal verification environments. Studying ISM-style memory in conjunction with dynamic in-model memory substrates also presents a compelling future direction.
Conclusion
ISM introduces a new standard for continual reasoning in memory-bounded, frozen-LLM environments by coupling schema-based strategy memory with autonomous self-improvement and symbolic validation. It establishes superior performance and memory efficiency on established mathematical reasoning benchmarks, providing both a practical toolkit and a direction for deeper integration between externalized, verifiable memory and modern LLM agents. ISM’s framework highlights the viability of small, actively maintained, and verifiable memory banks in supporting reliable cumulative learning beyond the limits of static parameterization.
